Multi-task Learning Long Short-term Memory Model to Emulate Wind Turbine Blade Dynamics

Abstract. The high computational costs in the dynamic analysis of wind turbines prohibit efficient design assessments and site-specific performance estimations. This study investigates the suitability of various dimensionality reduction techniques combined with a Long Short-term Memory (LSTM) algorithm to predict turbine responses, addressing computational challenges posed by high-dimensional inflow wind fields and complex time-stepping integration schemes. Feature selection criteria and a multi-stage modelling approach are implemented to arrive at a robust model configuration. Additionally, multi-task learning strategy is implemented which enables the LSTM model to predict multiple target variables simultaneously, eliminating the need for separate models for each target variable. Results demonstrate that this combined approach significantly reduces computational costs while maintaining consistent accuracy across all the target variables, thereby facilitating design feasibility studies and site-specific analyses of wind turbines.